Effects of Word Assignment in LDA for News Topic Discovery

In traditional LDA, latent variables are inferred from the "bag-of-words" assumption, in which word order is ignored. This bag-of-words assumption has gained recognition in terms of computational efficiency, whereas it is regarded impractical in many language model applications where word order is essential. In this study, we proposed word concatenation based on morphological rules as compounds and built the connection between compounds and topics. We used three categories including politics, economics, and life of Yahoo! Taiwan news from May/23/2013 to June/20/2013 and also extracted 1/3 of the news pool at random from each category as the mixed dataset. We compared unigrams and compounds in terms of topic coherence and performance, the result shows that the proposed model has a higher value of perplexity, while it illustrates more accurate meaning and computational efficiency than traditional LDA.